U.S. patent application number 14/258671 was filed with the patent office on 2014-08-07 for probability of accidental garment match.
This patent application is currently assigned to Quantum Signal, LLC. The applicant listed for this patent is Quantum Signal, LLC. Invention is credited to Norman H. Adams, Robert M. Lupa, Victor E. Perlin, Mitchell M. Rohde.
Application Number | 20140219513 14/258671 |
Document ID | / |
Family ID | 44278162 |
Filed Date | 2014-08-07 |
United States Patent
Application |
20140219513 |
Kind Code |
A1 |
Lupa; Robert M. ; et
al. |
August 7, 2014 |
Probability of accidental garment match
Abstract
A digital image of a first garment having one or more first
garment portions is received. A user has identified the first
garment portions as matching one or more corresponding second
garment portions of a second garment. The probability of accidental
match of the first garment within the digital image in relation to
the second garment is determined, by using a statistical model
based on one or more parameters and based on analyses of the first
garment portions. The probability of accidental match is
output.
Inventors: |
Lupa; Robert M.; (Vernon,
IL) ; Perlin; Victor E.; (Ann Arbor, MI) ;
Adams; Norman H.; (Fulton, MD) ; Rohde; Mitchell
M.; (Saline, MI) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Quantum Signal, LLC |
Ann Arbor |
MI |
US |
|
|
Assignee: |
Quantum Signal, LLC
Ann Arbor
MI
|
Family ID: |
44278162 |
Appl. No.: |
14/258671 |
Filed: |
April 22, 2014 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
12362482 |
Jan 29, 2009 |
8737741 |
|
|
14258671 |
|
|
|
|
Current U.S.
Class: |
382/111 |
Current CPC
Class: |
G06K 9/00362 20130101;
G06K 9/6262 20130101; G06K 9/6215 20130101; G06K 9/00771
20130101 |
Class at
Publication: |
382/111 |
International
Class: |
G06K 9/62 20060101
G06K009/62 |
Goverment Interests
GOVERNMENTAL RIGHTS IN THE INVENTION
[0001] The invention that is the subject of this patent application
was made with Government support under Contract No.
W91CRB-06-C-0013 awarded by the Department of Defense Combating
Terrorism Technology Support Office and the Technical Support
Working Group. The Government has certain rights in this invention.
Claims
1. A computer-readable medium having one or more computer programs
stored thereon for execution by one or more computing devices, such
that execution of the computer programs by the computing devices
cause performance of a method comprising: receiving a digital image
of a first garment having one or more first garment portions, a
user having identified the first garment portions as matching one
or more corresponding second garment portions of a second garment;
determining a probability of accidental match of the first garment
within the digital image in relation to the second garment, by
using a statistical model based on one or more parameters and based
on analyses of the first garment portions; and, outputting the
probability of accidental match.
2. The computer-readable medium of claim 1, wherein the digital
image is one of a still image and a frame of a video.
3. The computer-readable medium of claim 1, wherein the digital
image is one of surreptitiously taken and voluntarily taken of a
person wearing the first garment.
4. The computer-readable medium of claim 1, wherein the first
garment appears in the digital image within an environment, the
environment comprising one of a crime scene and a group photo.
5. The computer-readable medium of claim 1, wherein the second
garment is recovered from a person presumed to be a same person
wearing the first garment in the digital image.
6. The computer-readable medium of claim 1, wherein the digital
image is a first digital image, and the second garment appears in a
second digital image different than the first digital image, the
second digital image being one of a still image and a frame of a
video, the second digital image being one of surreptitiously taken
and voluntarily taken of a person wearing the second garment, the
second garment appearing in the digital image within an
environment, the environment comprising one of a group photo and a
controlled laboratory environment.
7. The computer-readable medium of claim 1, wherein the digital
image is taken from a crime scene in which a perpetrator of a crime
wore the first garment.
8. The computer-readable medium of claim 7, wherein the second
garment taken from a suspect of the crime.
9. The computer-readable medium of claim 1, wherein outputting the
probability of accidental match comprises displaying the
probability of accidental match on a display device.
10. The computer-readable medium of claim 1, wherein the user that
has identified the first garment portions as matching the second
garment portions is an expert user in garment matching.
11. The computer-readable medium of claim 1, wherein the first and
the second garments are made of a basic material having a detailed
repeating pattern.
12. The computer-readable medium of claim 1, wherein the first and
the second garments are camouflage uniforms.
13. The computer-readable medium of claim 1, wherein the first and
the second garments are military camouflage uniforms.
14. The computer-readable medium of claim 1, wherein the first and
the second garments are types of garments manufactured in multiple
factories according to at least substantially identical
specifications.
15. The computer-readable medium of claim 1, wherein the parameters
comprise a supposition that the garment including the first garment
portions and the garment including the second garment portions have
been manufactured using a same garment design template in that the
same garment design template has been used to guide cutting of raw
patterned fabric that was then sewn together to make the
garments.
16. The computer-readable medium of claim 1, wherein the parameters
comprise a supposition that the garment including the first garment
portions and the garment including the second garment portions have
been manufactured using different garment design templates in that
the different garment design templates have been used to guide
cutting of raw patterned fabric that was then sewn together to make
the garments.
17. The computer-readable medium of claim 1, wherein the parameters
comprise a supposition that the garment including the first garment
portions and the garment including the second garment portions have
been manufactured using face-to-face adjacent fabric plies within a
stack of fabric plies, in which pattern faces of the adjacent
fabric plies face one another, that are cut and then sewn together
to make the garments.
18. The computer-readable medium of claim 1, wherein the parameters
comprise a multiplicity of the garment including the second garment
portions in relation to the first garment portion, the multiplicity
denoting a number of areas of the garment including the second
garments portion that are visually similar to the first garment
portions, the areas including the second garment portions.
19. The computer-readable medium of claim 1, wherein the parameters
comprise a supposition that strict ply integrity has been
maintained during the manufacture of the garments, where strict ply
integrity means that at least medium and large pieces of a garment
are made from a common ply of fabric.
20. A system comprising: one or more computing devices; and,
hardware of the computing devices, the hardware programmed to:
receive a first digital image of a first camouflage garment; enable
a user to identify one or more first camouflage garment portions
within the first camouflage garment, the user having identified the
first camouflage garment portions as matching one or more
corresponding second camouflage garment portions of a second
camouflage garment; determine a probability of accidental match of
the first camouflage garment in relation to the second camouflage
garment, by using a statistical model based on one or more
parameters and based on analyses of the first camouflage garment
portions; and, output the probability of accidental match.
21. The system of claim 20, further comprising one or more digital
imaging devices by which the computing devices receive the first
digital image.
22. The system of claim 20, wherein the user is an expert user in
camouflage garment portion matching.
23. The system of claim 20, wherein the hardware is further
programmed to, after the first digital image has been received, and
after the user has identified the first camouflage garment
portions, for each first camouflage garment portion, blur a master
camouflage garment pattern to at least substantially match a
blurriness of the first camouflage garment portion within the first
digital image; determine one or more regions of the master
camouflage garment pattern that are visually similar to the first
camouflage garment portion; for each region of the master
camouflage garment pattern that has been determined, display the
region; permit a user to confirm visual similarity of the region to
the first camouflage garment portion.
24. The system of claim 20, wherein the parameters comprise a
multiplicity for each first camouflage garment portion, the
multiplicity equal in number to a number of the regions of the
master camouflage garment pattern that have been confirmed as being
visually similar to the first camouflage garment portion, plus a
number of the second camouflage garment portions corresponding to
the first camouflage garment portion.
25. The system of claim 20, wherein the hardware is to determine
the regions of the master camouflage garment pattern that are
visually similar to the first camouflage garment portion by
permitting a user to identify the regions.
26. The system of claim 20, wherein the hardware is to determine
the regions of the master camouflage garment pattern that are
visually similar to the first camouflage garment portion without
user interaction.
27. The system of claim 20, wherein the hardware is further
programmed to receive information regarding a manufacturing process
of at least the second camouflage garment, the information used in
determining a probability of accidental match of the first
camouflage garment in relation to the second camouflage
garment.
28. The system of claim 27, wherein the information comprises that
the first camouflage garment and the second camouflage garment have
at least likely been manufactured using a same camouflage garment
design template in that the same camouflage garment design template
has been used to guide cutting of raw camouflage fabric that was
then sewn together to make the first and the second camouflage
garments, such that the parameters comprise that the first
camouflage garment and the second camouflage garment have been
manufactured using the same camouflage garment design template.
29. The system of claim 27, wherein the information comprises that
the first camouflage garment and the second camouflage garment have
at least likely been manufactured using different camouflage
garment design templates in that the different camouflage garment
design templates have been used to guide cutting of raw camouflage
fabric that was then sewn together to make the first and the second
camouflage garments, such that the parameters comprise that the
first camouflage garment and the second camouflage garment have
been manufactured using different camouflage garment design
templates.
30. The system of claim 27, wherein the information comprises that
the first camouflage garment and the second camouflage garment have
at least likely been manufactured using face-to-face adjacent
fabric plies within a stack of fabric plies, in which camouflage
faces of the adjacent fabric plies face one another, that are cut
and then sewn together to make the first and the second camouflage
garments, such that the parameters comprise that the first
camouflage garment and the second camouflage garment have been
manufactured using face-to-face adjacent fabric plies within the
stack of fabric plies.
Description
BACKGROUND
[0002] Statistical evidence presents a unique challenge in legal
proceedings. Statistical evidence can provide weight for the
prosecution's claim that a given suspect has committed a particular
crime. Of particular concern here are instances in which a forensic
examiner, or expert, has established a match between a piece of
evidence found at the crime scene and a corresponding piece of
evidence directly related to the suspect him or herself--for
example, a crime scene image of a perpetrator in a unique set of
clothing, and similar if not identical clothing being found to
belong to the suspect. Even if such a match is incontrovertible,
the match may still be accidental.
[0003] Historically, a match is presented only if it could be
reasonably assumed that the probability of accidental match is
zero. For example, fingerprint evidence would only be admissible if
enough features are observed in the crime scene evidence to
preclude any individual other than the suspect having left the
fingerprints. However, modern forensic science has generally
recognized that such certainty can never truly be attained except
in a very small number of situations.
[0004] In some criminal investigations, critical pieces of evidence
include surveillance images showing the perpetrator of a crime
wearing clothing that can be matched to clothing worn or owned by a
suspect. Experts can, in some cases, find and document matching
areas between the clothing depicted in these images to the clothing
obtained from the suspect, and this evidence is specific enough so
as to constitute proof. These matches, however, can be imperfect
for a number of reasons, such as poor surveillance image quality,
inherent repetition of the garment manufacturing process leading to
garments having similar visual appearance, and so on. Thus, it is
important for forensic examiners, judges, juries, and others to
understand the quality of the match. Statistically analyzing these
garment matches can provide quantitative information on match
quality. Unfortunately, general garments are usually highly varied,
and in many cases not enough is known about them to estimate
meaningful statistics that can be used at trial.
SUMMARY
[0005] The present invention relates generally to determining the
probability of accidental match in garments, such as in the special
case of camouflage garments, and in particular military camouflage
uniforms. In criminal cases involving military personnel, the
widespread use of standard issue camouflage uniforms, such as the
US Army Combat Uniform (ACU), presents a unique context in which
statistical garment matching has found to be feasible. In this
case, a great deal is known about the garments and the people
wearing them. In particular, the garment manufacturing process has
been found to be sufficiently standardized that a statistical model
has been constructed that accounts for all significant sources of
variation within the garment. Embodiments of the present invention
employ a novel statistical model that is based on the knowledge of
the manufacturing process of camouflage uniforms, which is then
employed in conjunction with forensic data to estimate the
probability of an accidental camouflage uniform match.
[0006] A computer-readable medium of an embodiment of the invention
has one or more computer programs stored thereon for executing by
one or more computing devices. Execution of the computer programs
by the computing devices cause performance of a method. The method
receives a digital image of a first garment having one or more
first garment portions. A user has identified the first garment
portions as matching one or more corresponding second garment
portions of a second garment. The method determines the probability
of accidental match of the first garment within the digital image
in relation to the second garment, by using a statistical model
based on one or more parameters and based on analyses of the first
garment portions. The method outputs the probability of accidental
match.
[0007] A system of an embodiment of the invention includes one or
more computing devices that include hardware. The hardware is
programmed to receive a first digital image of a first camouflage
garment. The hardware is programmed to enable a user to identify
one or more first camouflage garment portions within the first
camouflage garment. The user has identified the first camouflage
garment portions as matching one or more corresponding second
camouflage garment portions of a second camouflage garment. The
hardware is programmed to determine a probability of accidental
match of the first camouflage garment in relation to the second
camouflage garment, by using a statistical model based on one or
more parameters and based on analyses of the first camouflage
garment portions. The hardware is further programmed to output the
probability of accidental match.
BRIEF DESCRIPTION OF THE DRAWINGS
[0008] The drawings referenced herein form a part of the
specification. Features shown in the drawing are meant as
illustrative of only some embodiments of the invention, and not of
all embodiments of the invention, unless otherwise explicitly
indicated, and implications to the contrary are otherwise not to be
made.
[0009] FIG. 1 is a flowchart of a method, according to an
embodiment of the invention.
[0010] FIGS. 2A and 2B are flowcharts of a method that is more
detailed than but consistent with the method of FIG. 1, according
to an embodiment of the invention.
[0011] FIG. 3 is a diagram of a computerized system, according to
an embodiment of the invention.
DETAILED DESCRIPTION OF THE DRAWINGS
[0012] In the following detailed description of exemplary
embodiments of the invention, reference is made to the accompanying
drawings that form a part hereof, and in which is shown by way of
illustration specific exemplary embodiments in which the invention
may be practiced. These embodiments are described in sufficient
detail to enable those skilled in the art to practice the
invention. Other embodiments may be utilized, and logical,
mechanical, and other changes may be made without departing from
the spirit or scope of the present invention. The following
detailed description is, therefore, not to be taken in a limiting
sense, and the scope of the present invention is defined only by
the appended claims.
Statistical Model for Determining Accidental Match Probability
[0013] In this section of the detailed description, the development
of a statistical model for determining the probability of an
accidental match between a first camouflage uniform and a second
camouflage uniform is described. The statistical model includes a
number of parameters. The camouflage uniforms from which the
uniform portions are selected or taken are exemplarily described
herein as US Army Combat Uniforms (ACU's), but other types of
camouflage uniforms are also amenable to embodiments of the
invention.
[0014] Let E.sub.C be the event that a surveillance image from a
crime scene that depicts a camouflage garment, or uniform, matches
a camouflage garment, or uniform, seized from a suspect. Matching
herein can mean that a forensic examiner, or expert or other user,
has established that these two pieces of evidence are compatible.
That is, there is agreement between all observable characteristics,
and there are no unaccounted-for differences. The likelihood ratio
for this match is:
LR = Pr ( E C | H 1 ) Pr ( E C | H 0 ) ( 1 ) ##EQU00001##
H.sub.0 and H.sub.1 are the two hypotheses. Specifically, H.sub.1
is the hypothesis that the garment seized from the suspect is the
garment depicted in the surveillance image. By comparison, H.sub.2
is the hypothesis that the garment seized from the suspect is not
the garment depicted in the surveillance image.
[0015] As such, Pr(E.sub.c|H.sub.0) is the probability, or
incidence rate, of a match occurring between the surveillance image
and a randomly chosen garment. Stated another way, this probability
is the accidental match probability. Addressed herein is the case
in which a forensic examiner has already established a match
between the two pieces of evidence in question, such that
Pr(E.sub.c|H.sub.1)=1. Accordingly, the likelihood ratio is the
inverse of the incidence rate--that is, the likelihood ratio is the
inverse of the accidental match probability.
[0016] To determine the accidental match probability--which can
include upper-bounding this probability in at least some
embodiments--a framework has been developed for piece-by-piece
matching of ACU garments. Each ACU garment is constructed from a
fixed number of pieces of fabric. The camouflage pattern is the
same for all fabric used in ACU garments, and repeats periodically
with a period L. Manufacturers of ACU garments may be required to
orient the fabric in one of two possible directions for all pieces
within the garment, but the portion of camouflage pattern that is
visible in each piece is not otherwise specified. As such, two
pieces of fabric with the same shape and orientation are identical
if a reference point in both pieces is located at the same position
in the L-by-L camouflage pattern.
[0017] However, in forensic applications just one garment is
physically available: the garment (or portions thereof) seized from
a suspect. The other garment is depicted in a surveillance image.
Surveillance images can be of low resolution, taken under poor
lighting conditions, may contain compression artifacts, and depict
garments that are wrapped around complex surfaces, such as the
human body. As such, the reference points from the same pieces of
the two garments may not need to come from exactly the same
location in the camouflage pattern for the two pieces to be judged
as matching. There exists a range of shifts for each reference
point within which a mismatch cannot be detected. The size of this
range is defined herein as the uncertainty of the match. This
uncertainty depends on the surveillance image, may be different for
each observable piece in the image, and must be determined by the
forensic examiner.
[0018] In addition, two unique pieces may be erroneously judged as
identical if they are drawn from distinct locations of the
camouflage pattern in which the pattern is similar. Due to the
algorithmic nature by which modern camouflage patterns (such as the
ACU) is generated, the pattern contains several areas that are very
similar, and that are virtually indistinguishable in low-quality
images, especially where just certain small areas of the pattern in
question are considered. Such similar portions of the pattern are
referred to herein as siblings. Some portions of the pattern have
many siblings, whereas others do not. For a given match between a
piece of the seized garment and a piece depicted in a surveillance
image, the number of siblings in the camouflage pattern that would
yield the same image is defined as the multiplicity of the match.
By definition, the match multiplicity is at least one, and every
portion of the camouflage pattern has at least one sibling--i.e.,
the portion itself. The match multiplicity depends on the size of
the observed portion of the pattern, and the image quality. As
such, the smaller the portion, the higher the multiplicity, and
likewise the poorer the image quality, the higher the
multiplicity.
[0019] Mathematically, a match between two pieces A and B is
expressed as:
|x.sub.A-x.sub.B-s.sub.x|<.DELTA.,|y.sub.A-y.sub.B-s.sub.y|<.DELTA-
., (2)
Here, (x.sub.A, y.sub.A) are the coordinates of the reference point
of piece A in the camouflage pattern, (x.sub.B, y.sub.B) are the
coordinates of the reference point in piece B, (s.sub.x, s.sub.y)
are the shifts between the true match location and a sibling, and A
is the match uncertainty. It is noted that (s.sub.x, s.sub.y)=(0,
0) for the match at the true match location. The coordinates
(x.sub.A, y.sub.A) and (X.sub.B, y.sub.B) are measured within a
single period of the camouflage pattern, and the difference
operation above operates on a circular space [0, L).
[0020] A statistical model for the variables in Equation (2) is now
developed. This model is based on the garment manufacturing
process, which is initially described. The model depends upon the
specific manufacturing parameters for each garment, and in
particular if the garments were manufactured from the same marker,
and if ply integrity was maintained for both garments.
[0021] The ACU garment manufacturing process may be divided into
four parts: marking, spreading, cutting, and sewing. Marking is the
process of creating a design template for cutting all of the pieces
needed for the garment from the camouflage fabric. This template is
referred to herein as a marker. Markers usually occupy the full
width of the fabric roll, minus a small margin on either side.
Markers are designed using manual layout software, automatic layout
software, or a combination of both. The principal goal in marker
design is to minimize the quantity of scrap fabric, without
complicating the cutting process unduly. A marker may be used only
once or many times over a period of years. The ACU garments are
made from a rip-stop material, as known within the art, and the
garment specifications restrict the possible rotations of the
pieces within the marker to either one specification orientation,
or a 180-degree rotation of that orientation.
[0022] Spreading is the process of laying plies of fabric on top of
one another. A stack of plies is referred to as a spread. Different
manufacturers use different spread thicknesses, from a few dozen to
a few hundred plies, as well as different spread lengths, from
several to a hundred yards. Typically, several hundred garments are
produced from a single spread. Spreading is accomplished either
manually or through automated spreading machines. Two methods of
spreading are common: single-sided and double-sided. Single-sided
spreads have all the plies face one direction (i.e., the camouflage
pattern face of each ply faces the same direction). Double-sided
spreads have pairs of plies that face each other (i.e., the
camouflage pattern faces of two adjacent plies face one another).
When spreading has been completed, a marker is placed on top of the
spread, and cutting begins.
[0023] Cutting can be performed either manually, by a skilled
laborer guiding a saw through the plies, or by an automated,
computerized cutting machine. In either case, the marker is used as
a guide for cutting out stacks of pieces. The garment
specifications may dictate that cuts have to be within a small
margin of error, such as 1/8 of an inch.
[0024] The stacks of pieces are handled with care as they move
through the sewing process. To maintain color shading consistency
across a given garment, specifications can require ply integrity,
which ensures that all the pieces used to make a given garment come
from the same ply in the spread. This is followed to varying
extents by garment manufacturers.
[0025] The stacks move through a factory-style assembly line of
sewing stations. Some sewing stations include a conventional sewing
machine and an operator, whereas others include elaborate
computer-controlled devices that perform one or more alignments,
stitches, and/or folds. Manufacturing facilities have varying
levels of automation, with some being almost entirely manual and
others being almost entirely automatic.
[0026] Piece-, or portion-, wise garment matches are determined
mathematically by the reference coordinates with respect to the
camouflage pattern. Therefore, a statistical characterization of
the coordinates is required to estimate the incident rate.
(x.sub.i, y.sub.i) are the longitudinal (along the spread) and
transverse (across the spread) coordinates of the position in the
camouflage pattern of a reference point for piece or portion i.
These coordinates can be expressed as:
x.sub.i=x.sub.M+x.sub.M,i+x.sub.cs,i (3)
y.sub.i=y.sub.M+y.sub.M,i+y.sub.cs,i
Here, (x.sub.M, y.sub.M) are the coordinates of the marker with
respect to the camouflage pattern, (x.sub.Mi, y.sub.Mi) are the
coordinates of the i-th piece within the marker, and (x.sub.cs,i,
y.sub.cs,i) are the combined cutting and stitching errors. The
cutting and stitching errors are defined as the deviations of the
reference position from the nominal due to imperfections in the
cutting and stitching processes. These errors are explicitly
limited to a small range by the garment specification, such as less
than 1/8 of an inch.
[0027] Each ply of a spread can begin at any point in the period L.
As such, the longitudinal coordinate of the marker x.sub.M is
unrestricted and can assume any value in the [0, L) interval with
uniform probability. By contrast, the transverse coordinate y.sub.M
ideally should be a constant for all plies in the spread. However,
in practice there are two factors that cause variation of y.sub.M
within a narrow range. First, the camouflage pattern tends to drift
transversely--i.e., the fabric edge corresponds to different points
of the pattern at different locations along the length of the roll.
Second, the fabric also tends to drift transversely as it is
spread--i.e., the edges of the spread fabric are not straight and
do not align perfectly from ply to ply. The former factor is
determined by the fabric manufacturer. The latter factor is
determining by the spreading process, and its magnitude varies
significantly between manufacturers.
[0028] Accurate statistical characterization of the coordinates
(x.sub.M, y.sub.M) is important within the statistical matching
model. As such, a method of direct sampling of spreads to estimate
the probability distribution functions (PDF's) of (x.sub.M,
y.sub.M) for manufacturers is developed. The method is based on
collecting and analyzing fabric pieces cut from the spread during
the manufacturing process. Every marker contains unused areas that
produce scrap stacks, which are typically discarded by the
manufacturer. For instance, numerous full scrap stacks may be
collected from several manufacturers. All pieces in each scrap
stack are then scanned, automatically aligned, and located within
the camouflage pattern based on the cross-correlation surface.
[0029] For all the pieces in a certain scrap stack, the second and
third terms of Equation (3) are the same, and the variation in
their coordinates is solely due to distribution of the marker
coordinates (x.sub.M, y.sub.M). Accordingly, given large enough
sample sets, the probability distributes for x.sub.M and y.sub.M
are found by constructing empirical histograms. Histograms for
individual stacks are combined to provide a single aggregate
histogram for each manufacturer. For the longitudinal direction,
individual stack histograms are combined directly, whereas for the
transverse direction the mean of each stack is subtracted prior to
combining the histograms. It has been found that the longitudinal
distribution is approximately uniform, and the transverse
distribution is approximately normal. The standard deviation of the
transverse distribution has been found to vary significantly
between manufacturers.
[0030] In addition to the marker coordinate distributions, the
dependence of the coordinates between plies is important for
garment matching. It has been found that, in general, plies are
independent, such that knowing (x.sub.M, y.sub.M) for a given ply
does not provide any information regarding another, randomly chosen
ply. However, this is not the case for adjacent plies, which is
relevant for double-sided spreading, in which the pieces of the
same garment often come from adjacent plies. Dependence between
adjacent plies is described by the condition PDF's,
Pr(x.sub.M,k|x.sub.M,k-1) and Pr(y.sub.M,k|y.sub.M,k-1).
[0031] For the longitudinal direction, the conditional distribution
is equivalent (same shape, different mean) to the distribution of
the difference (x.sub.M,k-x.sub.M,k-1), which is independent of
X.sub.M,k-1:
Pr(x.sub.M,k|x.sub.M,k-1).about.Pr(x.sub.M,k-x.sub.M,k-1|x.sub.M,k-1)=Pr-
(x.sub.M,k-x.sub.M,k-1) (4)
The PDF on the right side of Equation (4) is estimated from the
scrap stacks. If the fabric in the spread is laid continuously from
the same roll, the distribution acquires a Gaussian shape, as the
length of each ply is approximately constant. If there is a splice
point, due to a change of fabric rolls, or any other irregularity,
then the coordinates x.sub.M,k and X.sub.M,k-1 become decoupled,
which corresponds to uniform distribution tails. However, the
frequency of these irregularities is dependent upon varying
manufacturer practices, and is difficult to characterize. As such,
it is presumed herein that there are no such irregularities in the
spread, which narrows the distribution and increases the
probability of an accidental match:
Pr(x.sub.M,k|x.sub.M,k-1).about.N(.mu.,.sigma..sub.CL.sup.2),.mu.=x.sub.-
M,k-1+const (5)
[0032] Transverse marker coordinates in adjacent plies y.sub.M,k
and y.sub.M,k-1 are also highly dependent. Assuming that they
conform to a joint Gaussian distribution, the conditional PDF
Pr(y.sub.M,k|y.sub.M,k-1) is obtained, which is also Gaussian, with
non-zero mean and reduced variance:
Pr(y.sub.M,k|y.sub.M,k-1).about.N(.mu.,.sigma..sub.CT.sup.2),.mu.=y.sub.-
M,k-1r,.sigma..sub.CT.sup.2=.sigma..sub.T.sup.2(1-r.sup.2) (6)
[0033] Except for the term (c.sub.M, y.sub.M), all of the terms are
unique for each piece within a garment. By contrast, the term
(c.sub.M, y.sub.M) is often, but not always, constant for all
pieces within the garment. Indeed, (c.sub.M, y.sub.M) is the same
for all pieces within a garment if and only if ply integrity is
maintained for that garment. If two pieces come from different
plies, then each piece will have a different (c.sub.M, y.sub.M)
term. As such, knowledge of whether or not ply integrity was
maintained is important for determining the probability of an
accidental match, as each independent (c.sub.M, y.sub.M) term
reduces the probability by more than an order of magnitude. Poor
ply integrity greatly decreases the accidental match probability.
However, because an upper-bound to the probability of an accidental
match is sought by the overall estimation process, strict ply
integrity is presumed unless otherwise is known.
[0034] The cutting and stitching errors are difficult to measure
non-invasively within a real-world environment (wherein substantial
manual processes are used), such that empirical characterization is
nearly impossible. The garment specifications limit the range of
allowed cutting and stitching errors to about 1/8 of an inch each,
and it is reasonable to assume that these error distributions are
normal. As such, the cutting and stitching errors are combined into
a single Gaussian random variable with the standard deviation
supplied by the user. Forensic examiners typically will use a
conservative, small value, unless they have specific evidence to
the contrary. In all cases, the cutting and stitching error
distribution is narrow compared to the distributions that have been
described above, and has little influence on the incidence
rate.
[0035] The position of each piece within the marker
(x.sub.M,i,y.sub.M,i) is another potential source of variation, but
just in some situations. If the markers used to make the two
garments are independent, then the difference between the
corresponding (x.sub.M,i,y.sub.M,i) coordinates are uniformly
distributed on [0 . . . L).times.[0 . . . L), independently for
each piece. However, if the two garments are made from the same
marker, then the (x.sub.M,i,y.sub.M,i) coordinates are the same and
cancel one another in Equation (2). In this situation, if any one
piece matches between the two garments, and ply integrity was
maintained for both garments, then it is likely that all other
pieces will also match.
[0036] Between the boundary cases of completely independent markers
and identical markers is the case of similar markers, in which some
of the pieces have the same relative positions. The statistics of
similar markers are difficult to characterize, and such a
characterization would require access to a large number of sample
markers, which manufacturers are reluctant to release for
competitive reasons. To maintain a conservative upper bound on the
incidence rate, similar markers thus have to be treated as
identical.
[0037] The probability that two garments were made from the same
marker is difficult to estimate. ACU manufacturers continuously
discard and redesign their markers, while reusing some markers many
times. Hence, some markers may have been used for thousands of
garments, whereas others may have been used for only a few dozen
garments. Manufacturer records describing the use of markers are
not always kept, and thus cannot be assumed to be accessible or
even exist. Nevertheless, if the probability that the two garments
were made from the same marker was known to be P.sub.1, then it may
be incorporated into the incidence rate determination:
Pr(match)=Pr(match|identical markers)P.sub.1+Pr(match|independent
markers)(1-P.sub.1) (7)
[0038] In some cases it may be impossible to estimate P.sub.1, and
the forensic examiner must assume the worst case, that P.sub.1=1.
Overall, however, it has been found that it is unlikely that two
markers designed by different manufacturers are similar. This is
because different manufacturers use different spread lengths and
fabric widths. Many manufacturers design larger markers which
contain the pieces for multiple garments. In this case, the number
of possible arrangements becomes so large that the probability of
similar markers becomes negligible, even for just a few pieces. A
reasonable rule of thumb for upper bounding the incidence rate
(i.e., the accidental match probability) is to assume identical
markers for all garments made by the same manufacturer and
independent markers for garments made by different
manufacturers.
[0039] It is supposed that an n-piece match has been identified
between a garment depicted in a surveillance image (garment A) and
a garment seized from a suspect (garment B). An upper bound on the
probability that this match has occurred accidentally is sought.
The two sets of evidence are used to estimate a few parameters for
each piece i=1 . . . n, including the uncertainty A.sub.i, and the
multiplicity m.sub.i. The overall match is described as:
.A-inverted.i=1 . . . n.E-backward.j.epsilon.{1 . . .
m.sub.i}:|x.sub.i,A-x.sub.i,B-s.sub.x,i,j|<.DELTA..sub.i,|y.sub.i,A-y.-
sub.i,B-s.sub.y,i,j|<.DELTA..sub.i, (8)
[0040] It is noted that the match is separated into longitudinal
and transverse directions. This is tantamount to a square matching
area, with sides of length 2.DELTA..sub.i. Ideally, a circular
matching area with radius .DELTA..sub.i may be desirable, but this
would complicate the calculation significantly with little extra
knowledge gained. Because an upper bound to the match probability
is being determined, the larger square matching area is consistent
with the desired bound. The statistical characterization above
reveals that the longitudinal and transverse coordinates can
reasonably be modeled as independent, hence the probability
calculation is separable:
Pr(A and B match)=Pr(A and B match longitudinally)Pr(A and B match
transversely) (9)
[0041] For the probability calculation, the coordinates above are
broken into individual constituents, as in Equation (3), and
treated as random variables. The distributions of each random
variable are summarized as follows. The following difference random
variables are defined as:
.DELTA.x.sub.M=x.sub.M,A-x.sub.M,B,.DELTA.y.sub.M=y.sub.M,A-y.sub.M,B,
.DELTA.x.sub.M,i=x.sub.M,i,A-x.sub.M,i,B,.DELTA.y.sub.M=y.sub.M,i,A-y.su-
b.M,i,B,
.DELTA.x.sub.cs,i=x.sub.cs,i,A-x.sub.cs,i,B,.DELTA.y.sub.cs,i=y.sub.cs,i-
,A-y.sub.cs,i,B, (10)
Cutting and stitching errors are independent from one another, and
independent for each piece in each garment. Accordingly, if each is
modeled as Gaussian, then their linear combination is also
Gaussian:
.DELTA.x.sub.cs,i.about.N(0,.sigma..sub.cs.sup.2),.DELTA.y.sub.cs,i.abou-
t.N(0,.sigma..sub.cs.sup.2),.sigma..sub.cs.sup.2=.sigma..sub.cs,A.sup.2+.s-
igma..sub.cs,B.sup.2 (11)
[0042] Because the longitudinal coordinates are uniformly
distributed in the [0 . . . L) periodic pattern, their difference
A.sub.A/is also uniformly distributed over the same interval. The
transverse coordinates are normally distributed, hence their
difference is also normal, with double variance:
.DELTA.x.sub.M.about.U(0,L),.DELTA.y.sub.M.about.N(0,2.sigma..sub.T.sup.-
2) (12)
[0043] The piece coordinate differences within each marker
(.DELTA.x.sub.M,i,.DELTA.y.sub.M,i) are treated differently
depending upon whether the markers are assumed to be independent or
identical, where similar markers are treated as identical so as to
satisfy the upper bound. For the independent marker case, these
random variables are uniform on the interval [0 . . . L). For the
identical marker case, these terms become zero. The independent
marker case is first described now, followed by description of the
identical marker case.
[0044] If the markers used to make garments A and B are
independent, then the piece coordinate differences
(.DELTA.x.sub.M,i,.DELTA.y.sub.M,i) are also independently and
uniformly distributed over the pattern period. In this case, the
(.DELTA.x.sub.M,i,.DELTA.y.sub.M,i) random variables subsume the
other coordinate variables, and the overall coordinate differences
are independent and uniform:
.A-inverted.i=1 . . . n.A-inverted.j=1 . . .
m.sub.i:.DELTA.x.sub.i,AB=x.sub.i,A-x.sub.i,B=.DELTA.x.sub.M+.DELTA.x.sub-
.M,i+.DELTA.x.sub.cs,i.about.U(0,L),
.DELTA.y.sub.i,AB=y.sub.i,A-y.sub.i,B=.DELTA.y.sub.M+.DELTA.y.sub.M,i+.D-
ELTA.y.sub.cs,i.about.U(0,L) (13)
[0045] The probability of a match becomes the product of the
probability for each piece individually, and can be separated into
longitudinal and transverse components. The collection of m.sub.i
siblings where a match can occur are never close to one another
(within one to two inches), such that for each piece the m.sub.i
potential matches within the pattern are mutually exclusive events,
and their probabilities are simply summed. Thus, the probability
is:
Pr ( match | independent markers ) = Pr ( .A-inverted. i = 1 n
.E-backward. j .di-elect cons. { 1 m i } : .DELTA. x i , AB - s x ,
i , j < .DELTA. i , .DELTA. y i , AB - s y , i , j < .DELTA.
i ) = i = 1 n j = 1 m i Pr ( .DELTA. x i , AB - s x , i , j <
.DELTA. i ) 2 .DELTA. i / L Pr ( .DELTA. y i , AB - s y , i , j
< .DELTA. i ) 2 .DELTA. i / L = i = 1 n 4 m i .DELTA. i 2 L 2 (
14 ) ##EQU00002##
[0046] For typical values of the match uncertainty
.DELTA..sub.i.about.0.5 inches and pattern period L.about.34
inches, the probability of match for each piece is on the order of
.about.0.001. Hence, in the case of independent markers, the chance
of a match for more than one or two pieces is extremely low. By
contrast, match probabilities are not as low in the identical
marker case, which is now described.
[0047] If garments A and B were manufactured from the same marker,
then the piece coordinate differences (.DELTA.x.sub.M,i,
.DELTA.y.sub.M,i) are zero. Furthermore, for all pieces that come
from the same ply in the spread, the marker coordinate differences
(.DELTA.x.sub.M, .DELTA.y.sub.M) are identical, and the overall
coordinate differences for each piece (.DELTA.x.sub.i,
.DELTA.y.sub.i) are highly correlated. By contrast, each piece for
which ply integrity is not maintained has an independent
(.DELTA.x.sub.M, .DELTA.y.sub.M) term. Accordingly, the probability
of the overall garment match can be separated into the probability
that all pieces with ply integrity match and the probability that
each piece without ply integrity matches:
Pr ( all pieces match ) = Pr ( all PI pieces match ) i .di-elect
cons. NPI Pr ( i - th piece matches ) ( 15 ) ##EQU00003##
PI denotes the subset of pieces that are manufactured with ply
integrity, and NPI denotes the subset without ply integrity.
[0048] The pieces without ply integrity are considered first. Each
piece is independent of the others, and the longitudinal and
transverse directions for each piece are separable. Because
.DELTA.x.sub.M is uniformly distributed, the longitudinal component
of the probability is independent of s.sub.x,i,j. Using Bayes' Law,
the match probability is expressed as:
Pr ( i - th NPI piece matches ) = j = 1 m i Pr ( .DELTA. x M +
.DELTA. x cs , i - s x , i , j < .DELTA. i ) Pr ( .DELTA. y M +
.DELTA. y cs , i - s y , i , j < .DELTA. i ) j = 1 m i ( .intg.
Pr ( .DELTA. x M + .DELTA. x cs , i - s x , i , j < .DELTA. i |
.DELTA. x M ) Pr ( .DELTA. x M ) .DELTA. x M .intg. Pr ( .DELTA. y
M + .DELTA. y cs , i - s y , i , j < .DELTA. i | .DELTA. y M )
Pr ( .DELTA. y M ) .DELTA. y M ) = .intg. Pr ( .DELTA. x M +
.DELTA. x cs , i - s x , i , j < .DELTA. i | .DELTA. x M ) Pr (
.DELTA. x M ) .DELTA. x M j = 1 m i .intg. Pr ( .DELTA. y M +
.DELTA. y cs , i - s y , i , j < .DELTA. i | .DELTA. y M ) Pr (
.DELTA. y M ) .DELTA. y M ( 16 ) ##EQU00004##
[0049] To evaluate this integral,
Pr(|.DELTA.x.sub.M+.DELTA.x.sub.CS,I-s.sub.x,i,j|<.DELTA..sub.i|.DELTA-
.x.sub.M) is computed numerically by integrating over a portion of
a Gaussian curve (the PDF of .DELTA.x.sub.cs,i), or by using the
well-known Error Function:
Pr ( .DELTA. x M + .DELTA. x cs , i - s x , i , j < .DELTA. i |
.DELTA. x M ) = Pr ( - .DELTA. i - .DELTA. x M + s x , i , j <
.DELTA. x cs , i < .DELTA. i - .DELTA. x M + s x , i , j |
.DELTA. x M ) = CDF N ( .DELTA. i - .DELTA. x M + s x , i , j
.sigma. cs ) - CDF N ( - .DELTA. i - .DELTA. x M + s x , i , j
.sigma. cs ) ( 17 ) ##EQU00005##
This applies to the transverse direction as well.
[0050] A simple upper-bound may be used to reduce the number of
numerical integrations per NPI piece from m.sub.i to 1. Consider
the bottom integral over .DELTA.y.sub.M in Equation (16). The two
Pr( . . . ) terms under this integral are smooth, positive, and
symmetric functions of .DELTA.y.sub.M with single maxima: at
.DELTA.y.sub.M=s.sub.y,i,j for
Pr(|.DELTA.y.sub.M+.DELTA.y.sub.cs,i-s.sub.y,i,j|<.DELTA..sub.i|.DELTA-
.y.sub.M) and at .DELTA.y.sub.M=0 for Pr(.DELTA.y.sub.M).
[0051] Following this observation, the integral over .DELTA.y.sub.M
has the highest value when s.sub.y,i,j=0 and decreases
monotonically with increasing |s.sub.y,i,j|. This implies that each
spurious match (due to a sibling) contributes no more to the match
probability than the true match at (s.sub.x,i,j,s.sub.y,i,j)=(0,
0). With typical parameter values the probability becomes
negligible if s.sub.y,i,j exceeds one or two inches. The
probability of a match for the i-th NPI piece can therefore be
upper-bounded by computing the probability assuming that m.sub.i=1,
and then multiplying this probability by the number of siblings
within a small transverse shift
|s.sub.y,i,j|<2(.DELTA..sub.i+.sigma..sub.T+.sigma..sub.cs),
which is defined as the reduced multiplicity of the match
m.sub.i,0.ltoreq.m.sub.i. Therefore,
Pr(i-th NPI piece matches).ltoreq.m.sub.i,0Pr(i-th NPI piece
matches|m.sub.i=1) (18)
[0052] The pieces with ply integrity are now considered. Their
coordinate differences are dependent through the common marker
coordinate differences (.DELTA.x.sub.M,.DELTA.y.sub.M), but they
are conditionally independent given .DELTA.x.sub.M and
.DELTA.y.sub.M. As such, the probability of all PI pieces matching
is as follows:
Pr ( all PI pieces match ) = .intg. .intg. i .di-elect cons. PI Pr
( .E-backward. j .di-elect cons. { 1 m i } : .DELTA. x M + .DELTA.
x cs - s x , i , j < .DELTA. i .DELTA. y M + .DELTA. y cs - s y
, i , j < .DELTA. i | .DELTA. x M , .DELTA. y M ) Pr ( .DELTA. x
M ) Pr ( .DELTA. y M ) .DELTA. x M .DELTA. y M = .intg. .intg. i
.di-elect cons. PI ( j = 1 m i Pr ( .DELTA. x M + .DELTA. x cs - s
x , i , j < .DELTA. i .DELTA. y M + .DELTA. y cs - s y , i , j
< .DELTA. i | .DELTA. x M , .DELTA. y M ) ) Pr ( .DELTA. x M )
Pr ( .DELTA. y M ) .DELTA. x M .DELTA. y M ( 19 ) ##EQU00006##
[0053] The summation over m.sub.i in the bottom line of Equation
(19) can be replaced with a simple upper bound using a similar
argument to the NPI case described above, with one additional
constraint. Because all pieces with ply integrity share the same
marker shift, (.DELTA.x.sub.M,.DELTA.y.sub.M), all PI pieces match
compatible siblings simultaneously. The true match locations
guarantee at least one set of compatible siblings, at
s.sub.(x,y),i,j=1=(0,0) for all i. Any other set of siblings has to
satisfy .A-inverted.i,.E-backward.j.epsilon.{1 . . .
m.sub.i}:s.sub.x,i,j.apprxeq.s.sub.x. Accordingly, m.sub.0 is set
as the compatible multiplicity of the PI pieces:
.A-inverted.i.epsilon.PI.A-inverted.j.epsilon.(1 . . .
m.sub.0)|s.sub.x,i,j-s.sub.x,j|<2(.DELTA..sub.i+.sigma..sub.T+.sigma..-
sub.cs)I|s.sub.y,i,j|<2(.DELTA..sub.i+.sigma..sub.T+.sigma..sub.cs).
The probability that all PI pieces match is bounded by the product
of the match probability assuming the multiplicity of each piece is
one and the compatible multiplicity:
Pr(all PI pieces matches).ltoreq.m.sub.0Pr(all PI pieces
matches|.A-inverted.i.epsilon.PI:m.sub.i=1) (20)
[0054] Using the bound in Equation (20), the longitudinal and
transverse directions are again independent and the two-dimensional
integral in Equation (19) can be replaced with the product of two
one-dimensional integrals. In practice, m.sub.0 is virtually always
one if multiple PI pieces are observed. In this case, the
probability calculation is further simplified as:
Pr ( all PI pieces match ) = .intg. i .di-elect cons. PI Pr (
.DELTA. x M + .DELTA. x cs < .DELTA. i | .DELTA. x M ) Pr (
.DELTA. x M ) .DELTA. x M .intg. i .di-elect cons. PI Pr ( .DELTA.
y M + .DELTA. y cs < .DELTA. i | .DELTA. y M ) Pr ( .DELTA. y M
) .DELTA. y M ( 21 ) ##EQU00007##
[0055] The calculation that has been described above assumes
single-sided spreading. If double-sided spreading is employed, an
additional source of variation is introduced into the model and the
probability of accidental match may be reduced. The mode of
spreading affects the probability of accidental match only in the
case that ply integrity is maintained and identical markers are
used. Thus, the following description considers only PI pieces and
identical markers. Furthermore, because double-sided spreading is
only relevant in cases in which multiple PI pieces are observed,
the probability that m.sub.0>1 is negligible. The description
below is valid for m.sub.0=1. However, the approach can be
generalized for m.sub.0>1 if desired.
[0056] When double-sided spreading is used, the pieces of each
garment come from two adjacent plies in the spread, which are
referred to as the upper and lower plies. Accordingly, there are
now two pairs of marker shift differences instead of one pair:
(.DELTA.x.sub.M,u,.DELTA.y.sub.M,u) and
(.DELTA.x.sub.M,l,.DELTA.y.sub.M,l). These two pairs of differences
are highly correlated. Conditioning the upper ply differences on
the lower ply differences:
Pr(.DELTA.x.sub.M,u|.DELTA.x.sub.M,l).about.N(.DELTA.x.sub.M,l,2.sigma..-
sub.CL.sup.2)
Pr(.DELTA.y.sub.M,u|.DELTA.y.sub.M,l).about.N(.DELTA.y.sub.M,l,2.sigma..-
sub.CL.sup.2) (22)
[0057] Because m.sub.0=1, the longitudinal and transverse
directions can be separated. Considering just the longitudinal
component, the probability is:
Pr ( all pieces match ) = .intg. Pr ( all pieces match | .DELTA. x
M , l ) Pr ( .DELTA. x M , l ) .DELTA. x M , l = .intg. Pr ( upper
pieces match | .DELTA. x M , l ) Pr ( lower pieces match .DELTA. x
M , l ) Pr ( .DELTA. x M , l ) .DELTA. x M , l ( 23 )
##EQU00008##
Once the upper and low plies have been decoupled, the match
probability for the lower pieces is similar to the single-sided PI
case, and the match probability for the upper pieces is expressed
as:
Pr ( upper pieces match | .DELTA. x M , l ) = .intg. i .di-elect
cons. upper ply Pr ( .DELTA. x M , u + .DELTA. x cs , i <
.DELTA. i | .DELTA. x M , u , .DELTA. x M , l ) Pr ( .DELTA. x M ,
u | .DELTA. x M , l ) .DELTA. x M , u ( 24 ) ##EQU00009##
In this situation, a two-dimensional integral is needed, as the
shift differences have to be integrated for both the upper and
lower plies. However, numerical evaluation of such integrals is
relatively straightforward.
[0058] It is noted that in order to evaluate Equation (23), it has
to be known whether each piece came from the upper or lower ply.
Unless the specific marker used to manufacturer the garment is
available--which is unlikely--it is typically impossible to know
which ply each piece came from. If two mirror symmetric pieces are
observed, such as left and right trouser legs, then it may be
reasonable to assume that one comes from the upper ply and the
other from the lower ply. For other pieces, the ply is modeled as
an unknown variable, the probability determined for every valid
combination of upper and lower plies, and then averaged. Because
this iterative process may be slow, to maintain a valid upper-bound
on the match probability it may be preferred to assume that all
pieces come from the same ply, such that the probability reverse to
the single-sided case that has been described above.
Methods Employing Statistical Model for Determining Accidental
Match Probability
[0059] In this section of the detailed description, various methods
are described that employ a statistical model for determining
accidental match probability of a first camouflage uniform, such as
a US Army camouflage uniform (ACU), to a second camouflage uniform,
which also may be an ACU. The statistical model that is employed in
these methods may be the statistical model that has been described
in the previous section of the detailed description. Furthermore,
the methods may be implemented as or by one or more computer
programs that are stored on a computer-readable medium, such as a
tangible computer-readable medium like a recordable data storage
medium. The computer programs are thus executable by one or more
processors of one or more computing devices.
[0060] FIG. 1 shows a general method 100, according to an
embodiment of the invention. The method 100 may be implemented as
one or more computer programs that upon execution perform the
method 100. These computer programs may be stored on
computer-readable media, such as recordable data storage media.
[0061] A digital image of a first camouflage uniform is received
(102). For instance, the first digital image may be taken at a
crime scene in which a perpetrator of a crime wore the first
camouflage uniform. The first digital image is received from a
digital imaging device. For instance, the digital imaging device
may be a digital camera device that is capable of taking digital
photos, or a scanning device that is able to generate a digital
version of a film-based photo, as can be appreciated by those of
ordinary skill within the art. It may also be a still captured from
a digital or analog surveillance video or similar.
[0062] The first camouflage uniform includes one or more first
camouflage uniform portions, which a user has identified as
matching one or more corresponding second camouflage uniform
portions of a second camouflage uniform. The user may be an expert
user in camouflage uniform portion matching, such as a forensic
investigator. The second camouflage uniform is that which has been
taken from a suspect of the crime.
[0063] The method 100 determines a probability that the first
camouflage uniform accidentally matches the second camouflage
uniform (106). In one embodiment, determining this probability
specifically means that the upper-bound of the probability that the
first camouflage uniform accidentally matches the second camouflage
uniform. For example, if the upper-bound is 10%, then this means
that the probability that the first camouflage uniform accidentally
matches the second camouflage uniform is no greater than 10%, and
may be less. That there is a probability that the first camouflage
uniform accidentally matches the second camouflage uniform means
that, even though the user has matched the second camouflage
uniform portions to the first camouflage uniform portions, there is
nevertheless a probability that the camouflage uniform in the first
digital image is not the camouflage uniform seized from the
suspect.
[0064] The probability of accidental match is determined in part
106 of the method 100 by using a statistical model, such as the
statistical model described in the previous section of the detailed
description. Using the statistical model includes employing one or
more parameters and analyses of the identified first camouflage
uniform portions. Such parameters can include those that have been
at least implicitly referenced in the previous section of the
detailed description, for instance, and such analyses have also
been exemplary described above. However, some specific exemplary
parameters are now presented.
[0065] One such parameter is a supposition that the camouflage
uniform that includes the first camouflage uniform portion and the
camouflage uniform that includes the second camouflage uniform
portion have been manufactured using the same camouflage uniform
design template, or marker. That is, the same marker has been used
to guide cutting of raw camouflage fabric, which was then sewn
together to make the camouflage uniforms in question. This
parameter has been described in detail in the previous section of
the detailed description, and has been referenced as identical
markers.
[0066] By comparison, a second parameter is a supposition that the
camouflage uniform including the first camouflage uniform portion
and the camouflage uniform including the second camouflage uniform
portion have been manufactured using difference camouflage uniform
design templates, or markers. That is, different markers have been
used to guide cutting of raw camouflage fabric that was then sewn
together to make the camouflage uniforms in question. This
parameter also has been described in detail in the previous section
of the detailed description, and has been referenced as independent
markers. It is noted that the first parameter is mutually exclusive
with the second parameter, such that either the first parameter is
employed, or the second parameter is employed.
[0067] A third parameter is a supposition that the camouflage
uniform including the first camouflage uniform portion and the
camouflage uniform including the second camouflage uniform portion
have been manufactured using face-to-face adjacent fabric plies
within a stack of fabric plies. That is, the sides of two adjacent
fabric plies that have a camouflage pattern on them, which are
referred to as the camouflage faces of these plies, face one
another within the fabric ply stack. Thereafter, such fabric plies
are cut and sewn together to make the camouflage uniforms. This
parameter also has been described in the previous section of the
detailed description, and has been referenced as double-sided
spreading.
[0068] By comparison, a fourth parameter is a supposition that the
camouflage uniform including the first camouflage uniform portion
and the camouflage uniform including the second camouflage uniform
portion have been manufactured using fabric plies that have their
camouflage faces also oriented the same way (i.e., up or down)
within a stack of fabric plies. That is, the sides of two adjacent
fabric plies that have a camouflage pattern on them, which are
referred to as the camouflage faces of these plies, also face in
the same direction. Thereafter, such fabric plies are cut and sewn
together to make the camouflage uniforms. This parameter also has
been described in the previous section of the detailed description,
and has been referenced as single-sided spreading. It is noted that
the fourth parameter is mutually exclusive with the third
parameter, such that either the third parameter is employed, or the
fourth parameter is employed.
[0069] A fifth parameter is the multiplicity of the master pattern
from which the camouflage uniform including the second camouflage
uniform portion have been manufactured, where this multiplicity is
in relation to the first camouflage uniform portion. The
multiplicity denotes the number of areas of the master pattern of
the camouflage uniform, including the second camouflage uniform
portion, that are visually similar to the first camouflage uniform
portion. The multiplicity is at least one, because the areas
include the second camouflage uniform portion that has been matched
by the user to the first camouflage uniform portion. This parameter
also has been described in the previous section of the detailed
description, and has been referenced as siblings in addition to the
terminology multiplicity.
[0070] Once the probability of the accidental match has been
determined, it is output (108). For instance, the probability of
the accidental match may be displayed on a display device for
viewing by the user, or it may be printed on a hard medium like
paper by a printing device, also for viewing by the user. This
probability may be presented in the context of a report that
summarizes the input parameters, images, as well as other data.
Ultimately, the probability of accidental match may be used as
evidence that is presented within a criminal court proceeding in
which the suspect has been charged with being the perpetrator of
the crime (110). Presenting such accidental match probability is
used to at least assist in proving the guilt of the suspect in
committing the crime.
[0071] For example, a forensic investigator may be called by the
prosecution to testify that the suspect's camouflage uniform
matches the camouflage uniform taken in a surveillance photo at the
scene of the crime. To undercut the forensic investigator's
testimony, the defense may cross-examine the forensic investigator
and ask whether there is a chance that there is an accidental match
between the two uniforms. By employing an embodiment of the
invention, the forensic investigator is able to particularly
specify at least the upper-bound of such an accidental match. Where
the accidental match is relatively low, the jury or judge may be
convinced that there is no reasonable doubt that the garment worn
by an unknown individual in a surveillance image is the same as the
garment taken from a suspect's personal belongings. This lends
credence to the idea that the suspect is the one that committed the
crime, and should thus be convicted of it. By comparison, without
the invention, the forensic investigator is unlikely to be able to
quantitatively and particularly specify the probability of
accidental match, giving more credence to the defense's position
that there is reasonable doubt that the suspect has committed the
crime. This is particularly the case with camouflage uniforms,
which most laypeople tend to think of as "identical" though, with
sufficient resolution and study, one finds this to be untrue from a
scientific perspective.
[0072] FIGS. 2A and 2B show a method 200 that is more detailed than
but consistent with the method 100, according to an embodiment of
the invention. A first digital image of a first camouflage uniform
is received (202). The first digital image may be a digitized crime
scene surveillance photo in which a perpetrator of a crime is
wearing the first camouflage uniform.
[0073] A user is enabled or permitted to identify one or more first
camouflage uniform portions within the first camouflage uniform
(204). A user interface may be presented to the user of the first
digital image, for instance. As such, the user is permitted to draw
or otherwise select boundaries within the displayed first
camouflage uniform to select the first camouflage uniform portion.
As noted above, the user may be an expert user in camouflage
uniform portion matching, such as a forensic investigator. The
first camouflage uniform portions that the user identifies are
those that he or she believes match one or more corresponding
second uniform portions of a second camouflage uniform.
[0074] Thus, the user has identified the first camouflage uniform
portions as matching one or more corresponding second camouflage
uniform portions of the second camouflage uniform. The second
camouflage uniform may be that which has been seized from a suspect
of the crime. The user may, for instance, directly examine a
recovered uniform garment, or may examine an image--like a digital
image--of this second camouflage uniform. Such an image or images
may be created by photographing the recovered uniform garment in a
laboratory or other locale, for instance.
[0075] Information regarding the camouflage uniform manufacturing
process, of at least the second camouflage uniform, is received
(206). Such information is received to assist in later
determination of the accidental match probability. For instance,
the information can include that there is at least a likelihood
that the first and the second camouflage uniforms have been
manufactured from the same camouflage uniform design template, or
marker. Alternatively, the information can include that the first
and the second camouflage uniforms are unlikely to have been
manufactured from the same camouflage uniform design template, or
marker.
[0076] The information can further include that the camouflage
uniforms in question were at least likely manufactured using
double-sided spreading (i.e., where adjacent fabric plies within
the stack of plies have their camouflage faces facing one another).
Alternatively, the information can further include that the
camouflage uniforms were at least likely manufactured using
single-sided spreading (i.e., where all the fabric plies within the
stack of plies have their camouflage faces oriented in the same
direction). In general, the information received may include the
identification of the manufacturer of the second camouflage
uniform, where the additional information noted in this paragraph
is then looked up within a database of predetermined manufacturing
"profiles."
[0077] Thereafter, the method 200 performs the following for each
first camouflage uniform portion in question (208). The first
camouflage uniform portion is displayed for viewing by the user
(210). A digital image of a master camouflage uniform pattern is
then blurred (212). The master camouflage uniform pattern is the
pattern from which the second camouflage uniform, and presumably
from which the first camouflage uniform, were manufactured. The
master camouflage uniform pattern is leveraged by the model in
determining the probability of accidental match between the first
and the second camouflage uniforms. The insight here is to locate
within the master camouflage uniform pattern any other regions that
match the first camouflage uniform portions that the forensic
expert has already matched to the second camouflage uniform
portions--that is, other "sibling" areas as has been described
above. The greater the number of such additional master camouflage
uniform pattern regions that match the first camouflage uniform
portion in question, the greater the probability of accidental
match between the first and second camouflage uniforms.
[0078] Therefore, to assist the method 200 and/or the user in
locating additional regions within the master camouflage uniform
pattern that match the first camouflage uniform portion in
question, the master camouflage uniform pattern is artificially
blurred in part 212 to match the blurriness of the digital image of
the first camouflage uniform. For example, the user may be
permitted to select a blurring value on which basis the master
pattern is blurred and then displayed to the user. The blurred
master camouflage uniform pattern digital image may be displayed
alongside the first digital image (or the first camouflage uniform
portion that has been selected), so that the user can visually
assess whether the master camouflage uniform pattern has been
appropriately blurred. As such, the user can iteratively modify the
blurring value and then see the results of the blurring until the
user is satisfied that the master camouflage uniform pattern at
least substantially matches the blurriness of the first camouflage
uniform within the first digital image. The impetus for blurring
the master pattern is to ensure that, when searching for "sibling"
regions that also match the identified matching portions, the
increased number of (sibling) areas that appear visually identical
due to blur in surveillance photographs is appropriately determined
and passed on to the statistical model.
[0079] The method 200 then determines one or more regions (if any)
of the master camouflage uniform pattern that are visually similar
to the first camouflage uniform portion in question (214). It is
noted that this determination can be performed in one of two
different ways. First, the user him or herself may be permitted to
identify these additional regions. For instance, the user may be
permitted to manually identify these regions on the blurred master
camouflage uniform pattern using on-screen "paint" tools, where
such tools may also have been initially used to specify the
portions on the first digital image. Second, the method 200 itself
may analyze the master camouflage uniform pattern for regions
thereof that are visually similar to the first camouflage uniform
portion in question, using appropriate image-matching algorithms,
as can be appreciated by those of ordinary skill within the
art.
[0080] For each such additional region, the region is displayed,
and the user is permitted to confirm (or reject) the visual
similarity of the region to the first camouflage uniform portion
(216). For instance, the first camouflage uniform portion may be
displayed (as in part 210), and each additional region within the
master camouflage uniform pattern successively displayed next to
the first camouflage uniform portion in question. The user is then
permitted or enabled to approve the region currently being
displayed as being visually similar to the first camouflage uniform
portion in question, or to reject the visual similarity of the
region currently being displayed to the first camouflage uniform
portion in question. The user-approved "confirmed" regions are thus
the siblings, and the number of these additional visually similar
regions (plus one, for the base match) is equal to the
multiplicity.
[0081] Thereafter, the probability of accidental match of the
second camouflage uniform in relation to the first camouflage
uniform is determined by using a statistical model (106), based on
one or more parameters and based on analyses of the first and the
second camouflage uniform portions. As has been described, these
parameters can include whether there has been double-sided or
single-sided spreading, information regarding which has been
received in part 206. As has also been described, these parameters
can include whether identical markers or independent markers have
been employed, information regarding which has also been received
in part 206. As has also been described, the parameters can include
the multiplicity, which has been determined in part 208.
[0082] Once the probability of accidental match of the second
camouflage uniform in relation to the first camouflage uniform has
been determined, such as by using the statistical model described
in the preceding section of the detailed description, the
probability of accidental match is output (108). Such output can
include displaying the accidental match probability on a display
device, and/or printing this probability using a printing device.
Ultimately, the probability of accidental match can be used as
presented evidence within a criminal court proceeding (110), as has
been described above.
[0083] Embodiments of the invention are thus advantageous over the
prior art. While a user, such as a skilled forensic examiner,
inputs a qualitative match between camouflage uniform portions of
two digital images of camouflage uniforms, or that of a qualitative
match between camouflage uniform portions of one digital image and
one physical garment, embodiments of the invention output the
quantitative probability of accidental match. Thus information can
then be used for a variety of purposes, such as in legal
proceedings.
Computerized System
[0084] FIG. 3 shows a computerized system 300, according to an
embodiment of the invention. The system 300 includes one or more
digital imaging devices 302 and one or more computing devices 304.
The digital imaging devices 302 can include digital camera devices,
digital scanning devices, and other types of digital imaging
devices. The digital imaging devices 302 output a first image 306,
as has been described above. This image 306 is input into the
computing devices 304. That is, the digital imaging devices 302 are
communicatively coupled to the computing devices 304 in some
manner.
[0085] The computing devices 304 also can receive a digital image
of a master pattern 310 of the second camouflage uniform, as has
been described above. The computing devices 304 include hardware
312, such as memory, processors, storage devices, and so on. The
hardware 312 is specifically programmed to implement a model 316
for determining the probability of accidental match between the
first and the second camouflage uniforms. In this respect, the
hardware 312 is programmed to perform the methods that have been
described in the previous sections of the detailed description. The
methods and the model 316 may employ a database 314 that stores
information regarding how manufacturers makes camouflage uniforms,
to set one or more of the parameters of the model 316 that have
been described above. The methods provide the ability to input data
concerning the uniforms, images, manufacturer, and other salient
information.
[0086] On this basis, the computing devices 304 provide output 318.
The output 318 may be a printed report or information displayed on
a display device that is part of the hardware 312 of the computing
devices 304. The output 318 specifically includes the output of the
model 316, and thus the probability of accidental match between the
first and the second camouflage uniforms, as has been described
above.
CONCLUSION
[0087] It is noted that, although specific embodiments have been
illustrated and described herein, it will be appreciated by those
of ordinary skill in the art that any arrangement calculated to
achieve the same purpose may be substituted for the specific
embodiments shown. This application is intended to cover any
adaptations or variations of embodiments of the present invention.
Therefore, it is manifestly intended that this invention be limited
only by the claims and equivalents thereof.
[0088] For example, the camouflage uniform is more generally a
camouflage garment, which is itself more generally a garment. The
digital image in question may be a still image or a frame of a
video. The digital image may be surreptitiously taken from a person
wearing the first garment, or voluntarily taken (i.e., with
permission from the person). The first garment appears in the
digital image within an environment. This environment may be a
crime scene, a group photo, or another type of environment. The
digital image of the first garment may be taken from a crime scene
in which a perpetrator of a crime wore the first garment. As such,
the second garment may be taken from the suspect of the crime.
[0089] The second garment may be recovered from a person presumed
to be the same person wearing the first garment in the digital
image. The second garment may appear in a second digital image
different than the digital image of the first garment. This second
digital image may also be surreptitiously or voluntarily taken of
the person wearing the second garment, and the second garment may
appear in the digital image within an environment, such as a group
photo, or a controlled laboratory environment in which the second
garment is photographed after seizure from the owner (e.g., the
wearer) of the second garment.
* * * * *